venice-models

Files (1)

• SKILL.md

SKILL.md (first lines)

# Venice Models

Three read-only endpoints for model discovery - all `GET`, all public:

| Endpoint | Returns |
|---|---|
| `/models` | Model catalog with `model_spec` (capabilities, constraints, pricing). |
| `/models/traits` | Trait → model ID map (e.g. `default`, `default_reasoning`, `highest_quality`). |
| `/models/compatibility_mapping` | Legacy / third-party model ID → Venice model ID aliases. |

**Auth is optional.** These routes need no API key, so a plain `curl` works. If you send `Authorization: Bearer $VENICE_API_KEY` the result is tailored to that caller: a key's `modelPrivacy` setting filters the list, beta-flagged accounts also see beta models, and partner accounts see their negotiated rates. An invalid key is not rejected - you just get the public view.

**`?type=` values** (default `text` when omitted):

| Value | `/models` | `/models/traits` | `/models/compatibility_mapping` |
|---|---|---|---|
| `text`, `image`, `video`, `music`, `tts`, `asr`, `embedding`, `upscale`, `inpaint`, `decision` | yes | yes | yes |
| `all` (every type) | yes | yes | `400` |
| `code` (text models with `capabilities.optimizedForCode: true`) | yes | yes | `400` |

Anything else returns `400 { "error": "Invalid request parameters", "details": …, "issues": … }`.

`music` includes long-form audio (songs, sound effects, ElevenLabs narration models) and any voice-changer models. `decision` lists the typed-judgment models used by `POST /decisions` (today: `jev-latest`, flagged `betaModel: true`) - see [`venice-decisions`](../venice-decisions/SKILL.md).

## Use when

- You need to pick a model at runtime based on capabilities (vision, reasoning, function calling, E2EE, X search, multi-image, …).
- You need to validate a request against a model's `constraints` (prompt length, aspect ratio, resolution, steps, durations).
- You need the current **price per million tokens / per image / per second / per million characters** to build a cost estimate.
- You want to resolve a trait (`default`, `default_reasoning`, `highest_quality`) or a legacy alias (`gpt-4o`) to a concrete Venice model ID.

## `GET /models`

```bash
curl "https://api.venice.ai/api/v1/models?type=text"
```

```json
{
  "object": "list",
  "type": "text",
  "data": [
    {
      "id": "zai-org-glm-5-2",
      "object": "model",
      "owned_by": "venice.ai",
      "type": "text",
      "created": 1781568000,
      "context_length": 1000000,
      "model_spec": {
        "name": "GLM 5.2",
        "description": "GLM-5.2 is the next-generation large language model…",
        "availableContextTokens": 1000000,
        "maxCompletionTokens": 131072,
        "privacy": "private",
        "modelSource": "https://huggingface.co/zai-org/GLM-5.2",
        "offline": false,
        "traits": ["default", "function_calling_default"],
        "capabilities": { "supportsReasoning": true, "reasoningEffortOptions": ["none", "high", "max"], "…": "…" },
        "pricing": {
          "input":       { "usd": 1.4,  "diem": 1.4 },
          "cache_input": { "usd": 0.26, "diem": 0.26 },
          "output":      { "usd": 4.4,  "diem": 4.4 }
        }
      }
    }
  ]
}
```

### Top-level `ModelResponse` fields

| Field | Notes |
|---|---|
| `id` | The model ID to send as `model`. |
| `object` / `owned_by` | Always `"model"` / `"venice.ai"`. |
| `type` | One of the 10 model types above. |
| `created` | Unix seconds - release date on the Venice API. |
| `context_length` | Text models only. OpenAI-compatible mirror of `model_spec.availableContextTokens`. |
| `discount_to_user` | Reseller-only (0 < x < 1). Returned only to the reselling partner whose agreement it belongs to and omitted for other callers - treat absent as no discount. |
| `model_spec` | Everything below. |

### `model_spec` - common fields

| Field | Use |
|---|---|
| `name`, `description`, `modelSource` | Display name, blurb, upstream URL (`description` / `modelSource` may be absent). |
| `privacy` | `private` (zero data retention) or `anonymized` (third-party provider; not tied to your identity). |
| `offline` | `true` ⇒ requests return `503` "The model is temporarily offline". Skip it. |
| `traits` | Trait names this model currently holds (may be `[]`). |
| `uncensored` | Present and `true` only for models Venice classifies as uncensored (all modalities). Absent otherwise - never `false`. Upstream providers may still filter. |
| `betaModel` | Model is in beta status (still callable). |
| `beta` | Model is restricted to beta-access accounts. Only appears in lists returned to such accounts. |
| `regionRestrictions` | Country codes where the model is **blocked** (the OpenAPI description reads "intended to be available", but requests are rejected *from* the listed countries). Those requests get `403` "The specified model is unavailable in <country>…". Absent on unrestricted models. |
| `deprecation` | `{ autoRemap, date, removesAt, replacementModelId?, startsAt? }` - present only when retirement is scheduled. The model drops out of `/models` at `removesAt`; `autoRemap: true` means Venice may remap requests for this ID to `replacementModelId` instead of returning an error. |
| `model_sets` | Text, image and video only. Curation tags such as `venice_recommendations`, `featured`, and for video `audio`, `uncensored`, `high_resolution`, `fast`, … (served live but not declared in the OpenAPI schema). |

### `model_spec.capabilities` - text models

| Flag | Meaning |
|---|---|
| `optimizedForCode` | Tuned for coding tasks (drives `?type=code`). |
| `quantization` | `fp4` / `fp8` / `fp16` / `bf16` / `int8` / `int4` / `not-available`. |
| `supportsFunctionCalling` | `tools` are allowed. |
| `supportsResponseSchema` | Honors `response_format: { type: "json_schema" }`. |
| `supportsReasoning` | Model emits reasoning. |
| `supportsReasoningEffort` | Honors `reasoning_effort` / `reasoning.effort`. When `true`, also `reasoningEffortOptions` (subset of `none`, `minimal`, `low`, `medium`, `high`, `xhigh`, `max` - `none` means reasoning can be turned off) and `defaultReasoningEffort`. |
| `supportsVision` | Accepts `image_url` parts. |
| `supportsMultipleImages` + `maxImages` | More than one image per request; `maxImages` is the model's advertised limit. Chat hard-caps every model at 10 images per message. |
| `supportsVideoInput` + `maxVideos` | Accepts `video_url` parts; `maxVideos` present on some models. |
| `supportsAudioInput` | Accepts `input_audio` parts. |
| `supportsWebSearch` | `venice_parameters.enable_web_search` - currently `true` on every text model. |
| `supportsXSearch` | xAI native web + X search via `venice_parameters.enable_x_search`. |
| `supportsLogProbs` | Honors `logprobs` / `top_logprobs`. |
| `supportsTeeAttestation` | Runs in a TEE; verify with `GET /tee/attestation` / `/tee/signature`. |
| `supportsE2EE` | End-to-end encrypted inference (requires TEE). |

### `model_spec.constraints` and type-specific fields

From veniceai/skills, MIT.

More from veniceai/skills